Lorraine Weis, L3 Harris; Islam Hussein, L3 Harris
Keywords: Distributed Machine Learning, Space Situational Awareness, Recursive Neural Network, Sensor Tasking
Abstract:
Autonomous spacecraft swarms for Space Situational Awareness (SSA) promise to allow much greater coverage of the sky and supplement ground-based assets. By enabling swarm spacecraft to provide their own tasking and learn from past schedules, the swarm is much more flexible and resilient to changing environments. Autonomous tasking requires much less data be downlinked and can allow much faster responsiveness to shifting circumstances.
Space-based sensors for SSA have a number of advantages, that may offset the must higher cost of maintenance. First, they can be physically much closer to the objects of interest they are trying to measure. This enables more kinds of sensor technologies, such as laser ranging. Additionally, they are not subject to the atmospheric and geometric constraints that characterize ground based sensors. For example, a ground-based sensor can only observe a geostationary spacecraft at essentially one angle. Space-based sensors allow a much wider geometric diversity, which can greatly improve the observability of orbit parameters.
This work will investigate the effect of orbit design on the effectiveness of a space-based SSA swarm. Using a distributed deep machine learning tasking algorithm, swarms will be evaluated in terms of coverage and resiliency for a variety of target populations. Orbit diversity, number of swarm members, and other factors can greatly affect swarm performance in SSA.
Swarm sensor tasking for SSA is a many-to-many nonlinear assignment optimization problem. Fortunately, there are a number of simpler problems that are suggestive towards appropriate framing. For instance, the many-to-one linear tasking schedule problem can be framed as a convex optimization by relaxing the binary assignment to linear convex constraints. Furthermore, the orbital problem itself can be approximated as linear through coordinate choices such as Kustaaheimo-Stiefel coordinates, or by directly looking relative equinoctial elements or the Clohessy-Wiltshire-Hill equations.
For SSA decision making at the edge, we need algorithms that are lightweight in computation, memory, and communication. An exhaustive search of possible schedules is much too expensive to run on orbit, therefore lighter weight algorithms are needed. Furthermore, tasking strategies should be able to deal with a highly dynamic problem space – relative periodicities can shift substantially, targets may maneuver to new orbit parameters, unexpected events may strongly shift tasking priorities. Rather than having a hand tuned set of optimization heuristics, an autonomous SSA swarm should be able to use online training to adjust tasking schedules to fit the new environment, with minimal data transfer from the ground.
The distributed deep learning algorithm uses recursive neural nets to assign subsets of the target population to individual swarm members and then calculate near-optimal tasking schedules. While training may be initiated on ground-based processors, the intent is to allow continued online training on the spacecraft themselves, and use peer-to-peer communication to allow global optimization of the many-to-many tasking problem. The use of Recursive Neural Networks (RNNs) allows the space-based resources to maintain a simplified model of the nonlinear dynamics that can still adjust to changing environmental parameters. By using this underlying framework, we can evaluate the effectiveness swarm orbital designs directly, as well as gain intuition on which regimes such a swarm might be the most effective.
Date of Conference: September 17-20, 2019
Track: Space-Based Assets